Power quality 24-hour prediction using differential, deep and statistics machine learning based on weather data in an off-grid

Abstract

Prediction of Power Quality (PQ) on a daily basis is inevitable in planning Renewable Energy (RE) supply and scheduling the assumed power load in smart off-grid autonomous systems. Various combinations of the attached household appliances lead to specific operating states, related to the charge and load switching-time in different out-side conditions. Complexity and irregularities require modelling with the use of Artificial Intelligence (AI) to represent the uncertainty in load transitions and RE power oscillations. These specifics cannot be numerically solved with respect to unexpected states in detached micro-grid systems showing great variability. Recent Differential Learning, developed by the author, using a novel designed neuro-computing approach, enables one to model high non-linear and indefinable chaotic physical systems. The optimal training periods of day-set records were initially pre-assessed, based on AI input-output statistics. This model initialization allows operable day-ahead PQ-predictions in processing the last 24-hour series in one sequence. The proposed 2-level PQ-management evaluates initial load scheduling plans, based on available RE and storage sources. System efficiency and failure-free operations are planned considering the first estimate of its day-ahead PVP supply and secondary PQ verification of the provided load utilization schemes, under various state charges and RE production limits. This is a novelty with a notable incremental improvement in the published combinatorial optimization algorithms, only scheduling applicable load components, and reassessing the allowable power-day resources. Off-grid optimal operations can be determined and regulated by an adequate early morning PQ evaluation in RE utilization. A C++ parametric software, using differential learning to evolve PQ-models, historical PQ & meteo-data sets are at disposal to enable additional comparisons with the presented models.